All tools / MiniMax M3.1 Flash

Source-checked update

MiniMax M3.1 Flash: Open Weights and Deployment Choices

MiniMax M3.1 Flash weights are publicly available in the official Hugging Face repository. The release supports text, image and video input with text output. Choose a hosted service for a quick evaluation, or review the deployment recipes before downloading the weights.

Sources checked October 23, 2026 · No hands-on product test

What can you download?

The model card describes a mixture-of-experts model with approximately 428 billion total parameters, 23 billion active per token and up to a one-million-token context window. The repository includes a DSpark draft head. Its license is MiniMax Community; review the actual license before use or redistribution. Open weights do not mean unrestricted usage.

Inspect the model files · Read the model card

Choose your access route

Hosted evaluation

Use MiniMax Agent or the developer platform if you want to assess a task without provisioning inference hardware. Check account access, current charges and data handling on that service.

Self-hosted deployment

The model card points to SGLang, vLLM and TokenSpeed recipes. It lists several other integrations as in progress. Use the recipe for this exact release instead of assuming that every Transformers example already works.

Download only after planning the deployment

The model card provides this Hugging Face CLI command:

hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash

This downloads files; it does not start an inference server. First check storage, GPU memory and the serving implementation. Active parameter count is not a hardware memory estimate. The usable context and throughput depend on the configuration.

vLLM deployment recipe · SGLang deployment recipe

A useful first evaluation

  1. Pick one representative coding or document task and keep a copy of its expected result.
  2. Start with short input and record the exact model revision, serving version and settings.
  3. Check the result for correctness before measuring latency. Compare identical inputs and include failed requests in the result.
  4. Increase context only after measuring memory use. A stated maximum context is not evidence that your own deployment can serve it reliably.

We checked the public repository and documentation, but did not download the weight shards, run inference or reproduce benchmark scores.